Symbolic_Reasoning_Subsystem

Symbolic Reasoning Subsystem

brain/symbolic/ is the LLM-free reasoning core: world/causal/knowledge models plus a full rule lifecycle. Everything here runs in symbolic-only mode.

Module map

Reasoning and models

• inference.py — description-logic-style inheritance and queries over the world model.
• causal_graph.py — Pearl-style causal edges, structural updates, and intervention reasoning. The causal graph is a self-model: frontier goals derived from it point inward (search Orrin's own code and state), not at generic web research.
• prediction_engine.py — predictive-processing layer; prediction errors feed control signals.
• concept_formation.py / analogy_engine.py — forms new concepts, draws analogies across domains.
• reasoning_router.py — routes a question to the engine that can answer it.

The rule lifecycle

• rule_engine.py — evaluates rules each cycle; firings are logged.
• rule_synthesis.py — proposes new rules from observed regularities.
• rule_abstraction.py / rule_compressor.py — generalizes and compresses the rule base.
• rule_verifier.py — validates rules before they take effect.
• rule_forgetting.py — decays rules that stop earning their keep.
• meta_rules.py — rules about rules (seeded from brain/data/meta_rules.json).
• crystallization.py — hardens repeatedly successful reasoning into reusable skills.

Autonomy and measurement

• autonomous_experiment.py — designs and runs experiments to test its own hypotheses.
• intrinsic_motivation.py / progress_tracker.py / pattern_scorer.py — learning-progress signals that make reasoning self-directed.
• ground_truth.py / benchmark.py — checks conclusions against known facts; scores reasoning.
• self_improvement.py — turns recurring reasoning failures into improvement goals.
• llm_gate.py — the one place symbolic code may consult the LLM, subject to the global tool-only gate (see LLM Integration).

Production is visible

Symbolic productions — synthesized principles, crystallized skills, resolved experiments, established causal edges — are recorded as symbolic_artifact effects on the effect ledger via brain/symbolic/symbolic_effects.py, so symbolic work earns production reward the same way file outputs do (see Production and the Effect Ledger).

Persistence

World and causal models persist as JSON under brain/data/ and survive restarts; rule firings are logged to rule_firings.jsonl (capped).